Re: [libvirt] Using vmchannel?

I was just looking at the new vmchannel feature in QEMU and was wondering if we could make this available in libvirt. Firstly, since there isn't much documentation, I should explain how vmchannel works: [1] You pass an extra parameter on the qemu command line: qemu [...] -vmchannel <port>:<dev> where <port> is the TCP port number (see below) and <dev> is a standard qemu device description. As an example: qemu [...] -vmchannel 600:unix:/some/path [2] A new character device appears in the host, eg. Unix domain socket called "/some/path". [3] In the host, a userspace program should open a socket, bind(2) it to /some/path and listen(2) and accept(2) on it.

[4] In the guest, any process may connect(2) to TCP 10.0.2.4:600 (or whatever port was selected). Each connection in the guest causes the listener in the host to accept(2).

[5] This is only designed for low-bandwidth, low-performance applications. [6] Multiple vmchannels are supported. [7] Host cannot initiate a connection. My plan to implement this would be to add a new network interface type to the domain XML: <interface type='vmchannel'> <source port='600'/> <target path='/some/path'/> </interface> Only Unix domain socket paths would be allowed on the host side, and the path would be expected to exist already with suitable permissions. Note that I think this would also allow guests to communicate with the libvirtd on the host (not by default, of course, but if users wanted to configure it that way), for example: <interface type='vmchannel'> <source port='600'/> <target path='/var/run/libvirt/libvirt-sock'/> </interface> One problem is that it is qemu/kvm-only. It is built on top of virtio, and virtio is meant to become a standard driver subsystem for all full virtualization systems.